ShinrAI
VITAC IT-Systeme GmbH
Certified ShinrAI implementation and training partner.
Visit VITAC ↗
Semantic encryption
Keep the meaning. Protect personal data.
Detect personal data and replace it with consistent stand-ins before text reaches your AI workflow. Choose the managed API on STACKIT in Germany or deploy ShinrAI on your own infrastructure.
Multi-dimensional replacement
Bias-free by design
15 locales, one model
You hold the key
CPU or GPU — your hardware
Open weights — Innovius Open License
Privacy for business AI
ShinrAI on STACKIT
Business checkout, clear consumption tracking, API keys and invoices in one Dashboard. Monthly plans and non-expiring record packs.
1515 locales · Hebrew beta
STACKITGermany
€39€39 / month
2.8×
detection quality of a tuned rule engine on real German medical letters (F1 71.8 vs 25.6)
95,0
Median strict-span F1 · ShinrAI 1.3
<1ms
static screening path — regex, hash pools, gazetteers
15
locales served by the released ShinrAI 1.3 model
25ms
GPU inference per paragraph on a 2016 Tesla P40 (serve API, measured 2026-09) — L40S/A100 class: ≈7 ms (est.)
42
ranked replacement candidates scored per entity
Overview
What ShinrAI Does
Three levels of detail. Start where you want.
1
The one-sentence version
Before your text reaches any AI, ShinrAI swaps every person, place and company for a convincing stand-in — and swaps them back in the answer. The AI helps you without ever knowing who you are.
2
The practical version
A two-stage engine — a sub-millisecond static screen plus a purpose-trained neural detector — finds names, cities, streets and organizations in running text, across languages. Each finding is replaced by a context-matched counterpart, the AI answers naturally, and the reply is restored on your side. Masking is built in; replacement is why it works.
3
The engineering version
A 308M-parameter multilingual encoder (mmBERT base, ModernBERT family) with four entity heads and three attribute heads runs behind a regex/gazetteer stage. Replacement draws from realm-indexed pools — (type, origin, tier, gender, register) — with six validation gates and an injective session map. Restoration is a literal inverse over the stream; no server-side mapping store exists. Deployed as Helm charts, an OpenAI-compatible proxy, or a desktop connector.
Recorded synthetic example
See It in Action
The core idea
Why Replacement Beats Masking
Every classic tool can blank out a name. ShinrAI can too — but a blanked text loses the information the AI needs to reason. Multi-dimensional replacement keeps it.
Sample language:
Classic masking
Dear [PERSON], your appointment at [ORGANIZATION] in [CITY] on [DATE] is confirmed. Please bring [PERSON]'s referral from [PERSON].
- Which [PERSON] is which? Relations are destroyed.
- The AI sees broken, unnatural text — answers degrade.
- Gender, culture and formality vanish — or worse, get guessed.
- Still available in ShinrAI as mask mode when you want it.
ShinrAI semantic replacement
Dear Paul Brandt, your appointment at Klinikum Nordpark in Biberach on May 12 is confirmed. Please bring Anna Vogel's referral from Dr. Weber.
- Every stand-in keeps type, culture, gender, rarity and register.
- The AI reads natural text — reasoning quality survives.
- Identities stay distinct, so answers map back losslessly.
- A third mode, neutralize, strips bias signals on purpose.
| Original | Detected realm | mask | preserve | Why not the “obvious” swap? |
|---|---|---|---|---|
| Robert | PERSON · DE · common · masc · given | [FIRSTNAME] | Paul | Never “Mohammed” — flipping cultural origin measurably changes AI answers. |
| Bad Neustadt | CITY · DE · small | [CITY] | Biberach | Never “München” — a small-town→metropolis jump breaks the story’s logic. |
| Mailand (in German text) | CITY · IT · major | [CITY] | Turin | The entity is Italian even when the text is German — exonym-aware, culture kept. |
| Aleksandra Nowak | PERSON · PL · uncommon · fem · full | [PERSON] | Kasia Wiśniewska | Gender is never flipped; rarity tier is never dropped. |
| Siemens | ORG · DE · international | [COMPANY] | Bosch | An international brand stays an international brand — scale is meaning. |
| Autohaus Krüger GmbH | ORG · DE · regional | [COMPANY] | Fuhrmann Fahrzeugtechnik GmbH | Fictional but plausible — never a real firm, never a mismatched size. |
Worked examples from the ShinrAI replacement specification. Three policy modes — mask, preserve (default) and neutralize — are configurable per entity type and per deployment.
Multi-dimensional
Replacement in Five Dimensions
Each replacement matches the original in every dimension that shapes meaning — and passes six validation checks before it ships.
Type & tier
Person, city, street, organization — each with a rarity tier that is never dropped.
small town → small townCultural origin
50 origin classes. The entity’s culture is preserved — independent of the text’s language.
Mailand → Turin, not FrankfurtGender expression
Masculine, feminine, neutral. Flips are forbidden; fallbacks always go toward neutral.
Anna never becomes AndreasName part & register
Given, family or full name; formal, diminutive or abbreviated — the tone survives.
a nickname stays a nicknameSix validation gates
Distance ≥ 3 edits, length ratio bounds, mirrored capitalization, no collisions, injective map.
restoration is guaranteedAttribute detection measured at 97.7% (origin), 98.8% (gender expression) and 100.0% (name part) accuracy on the frozen v1-suite. The engine scores up to 42 candidates per entity; across all 24 language×type cells, at least 3 candidates land within 0.15 of the top score.
Architecture
How It Works
Two detection stages, one replacement layer, and a restoration path that only your side can compute.
Stage 1 — static screen
Regex patterns, O(1) name hash pools and city gazetteers catch structured PII — emails, phones, IBANs, tax and ID numbers — in under a millisecond, before any model runs.
10,000+ detection & replacement entries
Stage 2 — neural detection
A purpose-trained multilingual encoder finds what rules never can: names, places and organizations in running text, with tier, origin and gender attributes per span.
1024-token windows · sliding merge up to 128k docs
Lossless restore
The replacement map exists once — on your side. Answers stream back and every stand-in is swapped to reality mid-stream, even when a name splits across network chunks.
no server-side mapping store, by construction
Languages & cultures
Language Coverage
Identity looks different in every language and jurisdiction. ShinrAI models are trained per language track, with registry-anchored name, city, street and organization inventories per culture.
In production today
German
English
Japanese
French
Spanish
Italian
Polish
Portuguese (PT)
Portuguese (BR)
Russian
Ukrainian
Turkish
Arabic
Korean
Hebrew (beta)
All 15 locales run on the released ShinrAI 1.3 model — one model, automatic per-country conventions. ChtSafe and Secure AI Suite deployments run it in production; Hebrew ships as beta.
v1.4 — in training
Long documents
Paperwork register
Hebrew → full support
The 1.4 campaign trains at the Jülich Supercomputing Centre: long-document redaction, the paperwork register, and Hebrew toward full support — all current locales retrain alongside.
Specialist models — roadmap
Medical (ePA / clinical)
Coding & logs
Banking & finance
Government
Aerospace
Medical and aerospace lead the specialist queue; the rest follows the language tracks — the beta community votes on the order. Medical text is already a first-class training domain and our hardest benchmark.
Current model evidence
ShinrAI 1.3 is available across 15 locales, including Hebrew beta. ShinrAI 1.4 is in development, focusing on longer documents, paperwork and Hebrew. No release date announced.
95,0%
Median strict-span F1 · ShinrAI 1.3
Innovius evaluation: 200 business and clinical texts per locale, 15 locales, Hebrew beta included. Model-only scores; compare identical datasets and scoring methods.
Deployment
Deployment Options
The same protection, packaged for your cluster and your laptop.
ShinrAI Platform
Kubernetes-native engine with an OpenAI-compatible endpoint: drop it between your applications and any LLM. On-premise or VPC, CPU-only or GPU-served, air-gap ready. Streaming encryption and restore built in.
ShinrAI Connector
A desktop tray app that routes local AI tools — Claude Code, Codex, anything on the vendor SDKs — through on-device protection, and shows you exactly what it did. Monitor first, protect when ready.
Already protected products
ChtSafe brings ShinrAI-protected access to 50+ models for individuals. The Secure AI Suite deploys the full governed stack for organizations — ShinrAI is its encryption layer.
Enterprise add-on
Onion Routing: Hide Who Is Asking
ShinrAI protects what is in a request. Onion routing protects who sends it. Together they are the highest security level our enterprise customers run.
How it works
Requests travel to the AI provider over multiple independent routing hops and model access paths. No single hop sees your origin and your request together. The layer runs on llmproxy, the low-level C++ proxy engine behind the Secure AI Suite — built for high throughput at minimal added latency.
What it prevents
Your requests can no longer be attributed to your organization or person through provider billing accounts, IP addresses or country of origin. Competitors, providers and observers cannot tell that your company is asking — or from where.
How you get it
Optional, on top of any ShinrAI deployment. We demonstrate it live, and we plan, set up and operate it with your team as a paid service.
Request a demo & setup quoteContent protection (ShinrAI) works without onion routing. Add the routing layer when request attribution itself is a risk — trade secrets, M&A, defense, journalism, regulated research.
The question everyone asks
“What Do I Need to Run It?”
RAM and a fast CPU — or a GPU. Both work; memory is the main dimension. Honest numbers:
2 GB
RAM headroom for the compact INT8 model in process — measured ~1 GB resident after inference
4 GB
RAM minimum for the full-precision model — measured 2.0 GB resident per loaded session; 8 GB is comfortable
2–4
CPU cores per engine pod — inference threads pinned so the API never starves
any GPU
optional: even a 2016 Pascal card serves two models in ~4.5 GB VRAM; modern GPUs are effortless
Rule of thumb (measured 2026-09, fp32): an Apple M4 Pro detects a short message in 13 ms and a paragraph in 56 ms; a 4-vCPU x86 server VM needs 350 ms per paragraph; a 2016 Tesla P40 does it in 25 ms through the serve API — L40S/A100 class: ≈7 ms (est.). A Raspberry Pi 5 takes ≈982 ms per paragraph (est.) — batch use only. No internet access required at runtime.
Use cases
Where ShinrAI Is Used
Wherever people write about people, ShinrAI lets AI help without exposure.
Pharma & clinical research
De-identify electronic patient files (ePA), doctor letters and clinical archives on your own hardware — statistically faithful, analytically useful, never leaving the premises.
Real German doctor letters are the engine’s primary benchmark corpus. Batch API: 64 documents × 200k chars per call, horizontally scalable.
Banking, finance & insurance
Let advisors and analysts use frontier AI on customer correspondence while account holders stay unidentifiable — IBANs and card numbers caught in the static stage.
Structured identifiers (IBAN, cards, tax IDs) never depend on a model — the deterministic stage catches them in <1 ms.
Government & public sector
Citizen correspondence, case files and administrative records processed by AI inside your own security perimeter — with audit-grade, content-free logging.
Logs carry counts and types only. Config secrets are wiped from disk and env after load, into a memory-only store.
Legal
Drafting, summarizing and reviewing with AI on matters that name real people — while parties, addresses and firms travel as consistent, restorable stand-ins.
Session-consistent mapping: the same person keeps the same stand-in across the whole exchange, so cross-references survive.
Software teams
Route Claude Code, Codex and SDK-based tools through the desktop Connector — prompts are protected on-device before they leave the laptop, with a transcript of what happened.
Loopback-only listeners; a coding-specialist detection model is on the roadmap and beta testers vote on its priority.
Support & operations
Summarize tickets, draft replies and analyze feedback at scale — customers stay people to your team and patterns to the model.
Measured sustained throughput: hundreds of short requests per second on one CPU node; scales linearly with replicas.
Partners
Expertise from development to deployment.
Implementation and training partners
ELBA Technologies GmbH
Certified ShinrAI implementation and training partner.
Visit ELBA ↗Technology development partners & enablers
EECC
ShinrAI technology development partner.
Visit EECC ↗
Forschungszentrum Jülich
Research infrastructure and supercomputing enabler.
FZ Jülich ↗
Open research
Open Weights, Built in Public
Privacy infrastructure you can inspect beats privacy claims you must believe.
Open weights, fairly licensed
Every ShinrAI detection model ships as open weights on Hugging Face: v1.1 and v1.2 under Apache 2.0, the v1.3 flagship under the Innovius Open License — free for government, education, research, and companies under $10M revenue; every version becomes Apache 2.0 within 24 months.
Trained at Jülich
Developed with the EECC Research Lab and trained on the JURECA supercomputer at the Jülich Supercomputing Centre under the WestAI initiative — scaling onto JUPITER, Europe’s first exascale system.
open-weight generators only — no proprietary APIs anywhere
Anchored in open data
GeoNames, Wikidata, national statistics registries (SSA/Census, INSEE, INE, PESEL) give every culture honest, bias-aware ground truth — synthetic text, real distributions.
Open beta enrolling
Get Started
Join the model beta, bring the platform into your cluster, or start with the desktop Connector — and send the two-page overview to whoever asks what this is.
Support
How can we help?
Send a question about setup, billing or deployment. Include a request ID if useful; do not send API keys or personal data from your documents.
Phone supportNot available yet